From 8e4aae9339dcca0c49702f82acefb001d737470c Mon Sep 17 00:00:00 2001
From: Martin Bauer <martin.bauer@fau.de>
Date: Thu, 7 Mar 2019 16:07:04 +0100
Subject: [PATCH] CUDA backend: use fastmath by default
---
gpucuda/cudajit.py | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/gpucuda/cudajit.py b/gpucuda/cudajit.py
index ee40bd0c5..048557597 100644
--- a/gpucuda/cudajit.py
+++ b/gpucuda/cudajit.py
@@ -6,6 +6,9 @@ from pystencils.field import FieldType
from pystencils.include import get_pystencils_include_path
+USE_FAST_MATH = True
+
+
def make_python_function(kernel_function_node, argument_dict=None):
"""
Creates a kernel function from an abstract syntax tree which
@@ -33,8 +36,10 @@ def make_python_function(kernel_function_node, argument_dict=None):
code += "#define FUNC_PREFIX __global__\n"
code += "#define RESTRICT __restrict__\n\n"
code += str(generate_c(kernel_function_node, dialect='cuda'))
- mod = SourceModule(code, options=["-w", "-std=c++11", "-Wno-deprecated-gpu-targets"],
- include_dirs=[get_pystencils_include_path()])
+ options = options = ["-w", "-std=c++11", "-Wno-deprecated-gpu-targets", "-use_fast_math"]
+ if USE_FAST_MATH:
+ options.append("-use_fast_math")
+ mod = SourceModule(code, options=options, include_dirs=[get_pystencils_include_path()])
func = mod.get_function(kernel_function_node.function_name)
parameters = kernel_function_node.get_parameters()
--
GitLab